Dear researchers. I have created multiple clustering solutions for a group of objects by including (or not) diverse factors available, they determining the algorithms. Also, I have evaluated these clustering solutions and I got two measures of their "quality". My question is: how can I know the role of these factors in the obtention of good or bad evaluation measures? This is, is there a way of identifiying the positive or negative effect of including a factor in a clustering for the evaluation of this clustering? There are many clustering solutions and therefore ranking them according to the evaluation measures and looking at the features involved does not seem to be very informative. Many thanks!
http://www.ims.uni-stuttgart.de/institut/mitarbeiter/schulte/theses/phd/algorithm.pdf
good first reading
Hi Cesar,
In the field of Machine Learning is very common to use cross-validation for selecting the optimal parameters of a classifier/model. For example for selecting the hyperparameters of a neural network, regularization parameters, etc. So maybe you can do a similar procedure for the "factors" that you mention in your question.
Regarding the metrics, I will complement a bit the answer provided by Jesús. When you know the ground truth class assignments and the clustering assignments provided by your clustering algorithm, you can use the Mutual Information Based scores, such as Normalized Mutual Information (NMI) and Adjusted Mutual Information (AMI). In the case that the ground truth labels are not known, you can use the Silhouette Coefficient or the Calinski-Harabaz index. In the latter metrics, higher values of these coefficients imply a clustering algorithm that provides better defined clusters. You can review the implementation of the aforementioned metrics in scikit-learn (http://scikit-learn.org/stable/modules/clustering.html).
Kind regards.
Carlos.
- Badges
- Science topic
- Similar topics
- Comment